GPU-ENABLED FFANN INFERENCE FOR LOW-LATENCY VNF ORCHESTRATION IN MMWAVE 5G CORE NETWORK

Ultra-Reliable Low Latency Communication (URLLC) services in Millimeter-Wave Fifth Generation (mmWave 5G) core networks require efficient processing and real-time orchestration of network functions. Conventional Central Processing Unit (CPU)-based processing can introduce computational delays as telemetry volumes and network workloads increase. This study proposes a Graphics Processing Unit-enabled Feedforward Artificial Neural Network (GPU-FFANN) to predict Network Function (NF) overload and processing latency for proactive network orchestration. The proposed 11-64-32-2 FFANN architecture jointly performs NF overload classification and processing latency regression using a synthetically generated telemetry dataset of 300,000 samples, partitioned into 80% (240,000 samples) training, 10% (30,000 samples) validation, and 10% (30,000 Samples) testing. The model was implemented using PyTorch in a Computer Unified Device Architecture (CUDA)-enabled GPU computing environment. The GPU-FFANN performance was compared to a CPU/Virtual Central Processing Unit (VCPU)-based FFANN. The GPU-FFANN achieved 99.86% accuracy, 99.89% precision, 99.65% recall, and 99.77% F1-score for NF overload prediction, with a Mean Squared Error (MSE) of 0.2589 and a Mean Absolute Error (MAE) of 0.4055. Computational Inference performance was separately evaluated across inference batch sizes from 256 to 1,000,000 samples. At a batch size of 30,000, GPU-FFANN per-sample latency was approximately 5.0x10-8 s compared with 1.23x10-7 s for CPU/VCPU-FFANN, corresponding to a 59.35% reduction in per-sample inference latency and approximately 2.46 times faster inference. These results indicate that GPU-FFANN improves computational efficiency and supports responsive, proactive real-time VNF orchestration under high-traffic in mmWave 5G core networks.